Maybe [url=http://mersenneforum.org/showpost.php?p=228435&postcount=3]this post[/url] will help?

[quote=Karl M Johnson;229868]Maybe [URL="http://mersenneforum.org/showpost.php?p=228435&postcount=3"]this post[/URL] will help?[/quote]
That appears to be libcudart.so.2 (not 3); however, I'll give it a try and see if it works.

Edit: it didn't work.

I get the same error with tpsieve-cuda (see the Twin Prime Search forum), which uses libcudart.so.2, despite again having the file present in the tpsieve folder. So I'm wondering if it's something farther up the line, in my CUDA configuration.

Note also that /usr/local/cuda/lib is in my PATH, so it [i]should[/i] be able to grab the libraries from that folder. For whatever reason, though, it's not loading them even when I put the libraries right under its nose.

If you have installed the 64-bit CUDA version then the path should be /usr/local/cuda/lib64. I also needed to set the LD_LIBRARY_PATH variable under Linux. For Windows you just need to drop the dll into the same directory as the executable, but this didn't work for me under Linux.

[QUOTE=amphoria;229891]If you have installed the 64-bit CUDA version then the path should be /usr/local/cuda/lib64. I also needed to set the LD_LIBRARY_PATH variable under Linux. For Windows you just need to drop the dll into the same directory as the executable, but this didn't work for me under Linux.[/QUOTE]
Ah, I knew I was missing something simple like that...I remember one of the setup programs I ran to install the CUDA tools said something like that when it finished, but I was too trigger-happy and closed the console window before I got the chance to read it more fully. :rolleyes: I'll give it a try and see how it works.

Looks like it worked! :big grin: Now I have a new problem:
[code]gary@Buttford:~/Desktop/gpu-stuff/MacLucasFFTW$ ./MacLucasFFTW 21701
Error: API mismatch: the NVIDIA kernel module has version 71.86.04,
but this NVIDIA driver component has version 256.40. Please make
sure that the kernel module and all NVIDIA driver components
have the same version.
Error: API mismatch: the NVIDIA kernel module has version 71.86.04,
but this NVIDIA driver component has version 256.40. Please make
sure that the kernel module and all NVIDIA driver components
have the same version.
MacLucasFFTW.cu(960) : cudaSafeCall() Runtime API error : unspecified driver error.
[/code]
When I run "./MacLucasFFTW" by itself without an exponent, it prints the "API mismatch" error only once, then a few blank lines, and hangs there--pkill and Ctrl-C didn't work; I had to use -SIGKILL to stop it.

The drivers I have are the latest CUDA 3.1 developer toolkit. (3.2 beta required registration which I tried, but couldn't get to work.) The thing that stands out as strange to me here is that it's trying to compare "version 71.86.04" with "version 256.40"--they seem like two completely incompatible numbering schemes.

Note that prior to installing the 3.1 developer drivers, I had installed the "regular" latest GTX 460 drivers (I'm not sure what version they were). I got the impression, though, that installing the 3.1 developer drivers completely removed the original ones.

Meanwhile, now that the libcudart.so.3 issue has been resolved, I'll see if I have better luck with tpsieve. :smile:

I get the same error with tpsieve once it gets into the main calculative body of the program, so it looks like it's an all-around fatal error.

[QUOTE=mdettweiler;229910]I get the same error with tpsieve once it gets into the main calculative body of the program, so it looks like it's an all-around fatal error.[/QUOTE]

Hi, medettweiler
Ubuntu server is Difficult, Recommended Ubuntu Desktop.:lol:

[QUOTE=msft;229920]Hi, medettweiler
Ubuntu server is Difficult, Recommended Ubuntu Desktop.:lol:[/QUOTE]
Actually, this machine [i]is[/i] running Ubuntu Desktop. :smile:

Meanwhile, though, I figured out what the problem is. It seems that Ubuntu's restricted drivers manager has its own nVidia drivers it's trying to load, which are colliding with the ones I downloaded directly from nVidia. What I should have done is disable the restricted drivers manager [i]before[/i] installing the nVidia drivers--but it's too late for now. From what I've gathered around the internet, it seems that the only known solution for this is to reinstall the operating system (!).

However, in this case there's a possible better solution: Gary has three other computers that are absolutely identical to the GPU machine. I'll talk to him about moving the GPU to one of them, thus giving me a "clean slate" to work from. :smile:

Success! :big grin: Gary reinstalled the OS on buttford (he's got the process down to ~45 minutes by now--it was easier than moving the GPU to another machine along with the extra big power supply we bought to go with it), I got the drivers installed the right way today, and now I've just successfully ran my first test with MacLucasFFTW:
[code]M( 1257787 )P, n = 524288, MacLucasFFTW v8.1 Ballester[/code]
It seems that MLFFTW doesn't support FFT lengths <524288; because of this, the GPU took a lot longer to run this test than Prime95 did on my CPU since Prime95 could use the 64K FFT length. (Since MLFFTW only supports power-of-2 FFTs, I would have expected it to use n=65536 for this test. I'm not sure but that may even be the same as what Prime95 calls 64K.)

Congratulations, medettweiler
[QUOTE=mdettweiler;230024]It seems that MLFFTW doesn't support FFT lengths <524288;[/QUOTE]
Parallelization Side-effects.